Predicting CT-based coronal plane knee phenotype parameters using imageless navigation and machine learning.
Authors
Affiliations (3)
Affiliations (3)
- Sydney Orthopaedic Research Institute, Sydney, New South Wales, Australia. Electronic address: [email protected].
- Corin Group, Clinical Research, Raynham, MA, United States.
- Sydney Orthopaedic Research Institute, Sydney, New South Wales, Australia.
Abstract
Coronal plane alignment of the knee (CPAK) categorizes knee phenotypes according to joint line obliquity (JLO) and the arithmetic hip‒knee‒ankle angle (aHKA). CPAK is traditionally measured via long leg radiographs, but recently other modalities such as computed tomography (CT), image-based, and imageless navigation have been used. Machine learning (ML) is a field of artificial intelligence focused on enabling systems to learn from data and improve their performance without explicit programming. The aim of this study is to understand how accurately imageless navigation measures CPAK parameters relative to CT using generic cartilage wear assumptions and ML models. 152 TKAs performed via imageless navigation with preoperative CT data were retrospectively reviewed. MPTA and LDFA were measured from both the preoperative CT and the intraoperative imageless navigation landmark data and used to calculate JLO and aHKA. Three articular cartilage wear assumptions were applied to the imageless navigation data. The first applied no wear correction, whilst the second used traditional preoperative coronal HKA thresholds with a 2 mm cartilage thickness wear assumption on both the femoral and tibial side. A third model used retrospective thresholds aimed at optimally reducing error. Mean, SD, mean difference and Mean Absolute Error (MAE) were calculated. A fourth ML model, using random forest modeling with cross validation on an 80-20 test-train split, was used to determine MAE of navigated data from corresponding CT values. The average age was 73 ± 8 years, with 61% women, and an average preoperative coronal deformity of 3.5 ± 4.2° varus. The ML based wear assumptions had the lowest MAE for all CPAK parameters, with MAE ≤1.2° for MPTA and LDFA, and ≤1.8° for JLO and aHKA. This was better than MAE for aHKA for the no wear assumption model (2.5°) and generic wear model (2.6°). Imageless navigation can measure MPTA and LDFA with a mean error of <1.2° compared to CT when using ML models to predict cartilage wear. These results indicate that imageless navigation can be used to effectively measure CPAK parameters, achieving comparable results to a CT-based approach.